structured knowledge and novel object kinds can be …...main effect of condition f(1,47) = 12.17, p...
TRANSCRIPT
University of PennsylvaniaAnna Leshinskaya & Sharon L. Thompson-Schill
Gopnik, A., & Sobel, D. M. (2000). Detecting blickets: how young children use information about novel causal powers in categorization and induction. Child Development, 71(5), 1205–1222. Kemp, C., Tenenbaum, J. B., Niyogi, S., & Griffiths, T. L. (2010). A probabilistic model of theory formation. Cognition, 114(2), 165–96.
Structured knowledge and novel object kinds can be inferred from visual event streams
One way to achieve abstraction from sensory experiences is to encode relations among sensory events; a basic one being the direction of prediction ( ‘before’ vs ‘after’).
To what extent do adult learners do this spontaneously in naturalistic and bottom-up fashion?
Statistical information should spontaneously inform con-ceptual judgment and category formation without a top-down instructional context.
Relational schemas should operate fairly automatically during event processing.
e�ect
cause
frequent
rare
static
.02
.43
.53.02
.36
.35
.25
.04.70.01
.10
.04
.02
.92
.05.02
.02
.49
.01
.25
.25
.15
e�ect
cause
rare
static
.02
.43
.53
.36
.35
.25
.04.70.01
.10
.04
.02
.93
.05.02
.02
.01
.25
.25
.15frequent
thale sibbie
causality ratings
Task: pay close attention and try to predict what will happen next (press a key when something unexpected occurs).
2.4 s 1.2 s
Stimuli: sequence of 250 animated visual events for 2 objects, order governed by markov chain.
Experiment 1: How objects attain causal properties from naturalistic event streams
Experiments 2 & 3: Relational schemas are deployed automatically
Experiment 4: Unsupervised induction of causal kinds, generalizing across manner of
causation
cause
effect
light-causers light-reactors
cause
effect
cause
effect
cause
effect
shape to condition assignments and effect events counterbalanced
training objects
1 2 1 2
task same as Experiment 1 with one interim non-specific test
similarity judgmentsfollowing training using a spatial sorting task
Linear regression with motion model : c cells = 0 , others = ; causal model : m cells = 0 , others = 1, fit on individual data, betas subjected to t-test.
Linear regression with a single factor:mixture model: c cells = .5 and m cells = .5, others = 1
more sim
ilarm
ore di�erent
light causer 1
light causer 2
light reactor 1
light reactor 2
light causer 1light causer 2light reactor 1light reactor 2
C
C
m
m
generalization testwhich objects is this most similar to?
same motion & same causality
different motion & same causality
same motion & different causality
different motion & different causality
cause+ motion+
cause+ motion-
cause- motion+
cause- motion-
0 0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Percent of Subjects Choosing
n=18
Motion match F(1,17) = 15.36, p < .001Causal match F(1,17) = 5.52, p < .05
Casual modelt(17) = 2.42, p < .05
Motion model t(17) = 2.39, p <.05
Mixture model t(17) = 3.53, p < .01
rare
common
common
common
rare
response
1.2s
oddball decision task
obj
ect 1
obj
ect 2
different-relation condition
same-relation condition
shape to condition assignments randomizedeffect event (light, bubbles, confetti) counterbalanced
order of training objects randomized
familiarity forced choice teststrong transitions compared to weaker transitions
administered following all learning exposure
which of the two videos is more typical/familiar?
oddball events on 10% of trials
obj
ect 1
obj
ect 2
different-relation condition
same-relation condition
obj1 obj2 obj1 obj20
0.10.20.30.40.50.60.70.8 ** **
*
different-relation same-relation
cause to e�ectThales tilting causes the light to flash.
e�ect to causeThe light flashing causes thales to tilt.
object to e�ectThales cause the light to flash.
object to frequentThales cause the stars to appear.
1 2 3 4 5definitely false definitely trueunsure
1
1.5
2
2.5
3
3.5
4
4.5
5
caus
e-e�
ect
e�ec
t-cau
se
obje
ct-e
�ect
obje
ct-fr
eque
nt
***
** *****
**
**
n = 57n = 24
n = 14
** p <.01 *** p< .001
Main effect of condition F(1,70) = 5.14, p < .05
- Higher (40%) noticing rate due to slight changes in the task. Thus, par-ticipants split into noticers and non-noticers using freeform responses post-task.
-Training object 2 was always tested first.
-Same-relation condition: ‘cause’ event was ambient for both videos, or object-based for both videos.
Subjects did not have explicit access to this knowledge.
object 1 object 2 object 1 object 20
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
same-relation different-relation
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
object 1 object 2 object 1 object 2
same-relation different-relation
noticers non-noticers
Main effect of condition F(1,47) = 12.17, p = .001
Main effect of condition F(1,82) = 8.00 p = .006
Noticing by Condition interaction
F(1,129) = 17.80, p < .0001
It should be noted that we did not see this effect in a previ-ous version of this
task (with some dif-ferences) and thus aim to replicate it.
**
**
familiarity forced choice test
(by training order)
oddball decision task
Why is this important?
Predictive structure is a pervasive part of experience that can be extracted using straightforward learning mechanisms. But it can also be leveraged to gain abstraction, as predictive relations can be generalized across participating events and sensory features. Together, this could account for bottom-up abstraction of sen-sory experience and the formation of novel kinds generalizing across sensory features. do
wnl
oad
me!
normalized screen distances
compared only in participants accurate on non-causal judgments of the relevant statistic